A deep learning algorithm to detect proximal humerus fractures on radiographs.

Area of Science:

• Radiology
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Proximal humerus fractures are common in the elderly, with inconsistent classification by traditional methods.
• Poor agreement exists in classifying these fractures using current schemes.
• Deep learning (DL) shows potential for improving fracture pattern recognition.

Purpose of the Study:

• To develop a reliable deep learning (DL) approach for detecting proximal humerus fractures on radiographs.
• To ensure DL fracture classifiers are applied only to X-rays with actual fractures.

Main Methods:

• Trained a DL model on 996 fractured and 607 non-fractured proximal humerus radiographs.
• Utilized five-fold cross-validation for model development and an internal test set.
• Validated the best model on an external test set of 116 radiographs.

Main Results:

• The DL model achieved 0.972 accuracy and 0.969 F1 score on the internal test set.
• External validation showed 0.966 accuracy/sensitivity, misclassifying only 4 fractures.
• Saliency maps indicated the model focused on the humeral head's perimeter for detection.

Conclusions:

• The developed DL algorithm reliably detects proximal humerus fractures on radiographs.
• This tool can assist emergency departments by flagging potential fractures for review.
• It serves as a foundation for future AI tools in managing these fractures.